The Information Test

To test whether tax misperceptions can alter labor supply, Dur, Harms, and Non had to change information without changing the tax system itself. This chapter follows their Dutch experiment from the first report of beliefs and work hours, through random assignment and the information package, to the second survey two months later. The design matters because it tells us what the later comparison can credibly mean—and why an offer of information, especially an unused personalized tool, is not the same as a precise tax calculation delivered and acted upon.

Before the information: establishing a baseline

The study used two waves of a survey conducted through the LISS panel in August and October 2024. The first wave did more than ask respondents whether they thought taxes were high. It elicited a concrete belief about the effective marginal return: how much the respondent would gain on net if gross monthly income increased by €100. The survey also recorded the respondent’s current weekly work hours, preferences over job aspects, and whether particular circumstances limited the possibility of increasing hours.

That combination of measures gives the experiment a starting state. The belief says what the respondent thinks an additional increment is worth. Hours show the current labor-supply position. The questions about health, childcare, caregiving, and study indicate whether an hours response is feasible. Without the belief measure, a later change in hours could not be connected to the proposed perception mechanism. Without hours, a belief correction could not be related to behavior. Without constraints, a failure to change hours could be mistakenly read as evidence that the information had no relevance.

Respondents entered a common measurement stage before the information was supplied. They typed a number for the expected net gain and then rated how certain they were about that answer. The open response allowed them to express a belief that the marginal rate was unusually high or low, including a negative number for the net change if they thought an income increase would reduce disposable income. The study later excluded extreme beliefs implying a negative tax rate or a rate above 200%, but the purpose of the question was to preserve the range of actual uncertainty rather than force every answer into a familiar tax bracket.

The sample became smaller at each transition. Of 4,859 eligible LISS-panel members invited to the first wave, 3,274 completed it, a response rate of 67.4%. People over 67 and those unable to work because of disability were not invited, reflecting the study’s focus on labor supply. All first-wave respondents were approached again in October; 2,788 completed the second wave, which the paper reports as an 83.5% response rate. After incomplete information and missing income reduced the usable paired responses to 2,587, 46 respondents with extreme beliefs were removed. The main analysis then excluded 663 people assigned to the personalized-information arm, leaving a final analytical sample of 1,878.

These figures are not a criticism of the design by themselves. A study that follows people over time will lose some respondents, and the researchers report that attrition was about 23% and similar in treatment and control groups. They also report no statistically significant difference at the 5% level in the baseline characteristics shown in the treatment and control groups. Together, those facts support the use of the randomized comparison.